A computationally efficient speech/music discriminator for radio recordings

This paper presents a speech/music discriminator for radio recordings, based on a new and computationally efficient region growing technique, that bears its origins in the field of image segmentation. The proposed scheme operates on a single feature, a variant of the spectral entropy, which is extracted from the audio recording by means of a short-term processing technique. The proposed method has been tested on recordings from radio stations broadcasting over the Internet and, despite its simplicity, has proved to yield performance results comparable to more sophisticated approaches.